1 //===- LoopRotation.cpp - Loop Rotation Pass ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements Loop Rotation Pass.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "loop-rotate"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/Function.h"
17 #include "llvm/Analysis/CodeMetrics.h"
18 #include "llvm/Analysis/DominanceFrontier.h"
19 #include "llvm/Analysis/LoopPass.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/ScalarEvolution.h"
22 #include "llvm/Transforms/Utils/Local.h"
23 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
24 #include "llvm/Transforms/Utils/SSAUpdater.h"
25 #include "llvm/Transforms/Utils/ValueMapper.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/ADT/Statistic.h"
30 #define MAX_HEADER_SIZE 16
32 STATISTIC(NumRotated, "Number of loops rotated");
35 class LoopRotate : public LoopPass {
37 static char ID; // Pass ID, replacement for typeid
38 LoopRotate() : LoopPass(ID) {
39 initializeLoopRotatePass(*PassRegistry::getPassRegistry());
42 // Rotate Loop L as many times as possible. Return true if
43 // loop is rotated at least once.
44 bool runOnLoop(Loop *L, LPPassManager &LPM);
46 // LCSSA form makes instruction renaming easier.
47 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
48 AU.addPreserved<DominatorTree>();
49 AU.addPreserved<DominanceFrontier>();
50 AU.addRequired<LoopInfo>();
51 AU.addPreserved<LoopInfo>();
52 AU.addRequiredID(LoopSimplifyID);
53 AU.addPreservedID(LoopSimplifyID);
54 AU.addRequiredID(LCSSAID);
55 AU.addPreservedID(LCSSAID);
56 AU.addPreserved<ScalarEvolution>();
62 bool rotateLoop(Loop *L);
64 /// After loop rotation, loop pre-header has multiple sucessors.
65 /// Insert one forwarding basic block to ensure that loop pre-header
66 /// has only one successor.
67 void preserveCanonicalLoopForm(Loop *L, BasicBlock *OrigHeader,
68 BasicBlock *OrigPreHeader,
69 BasicBlock *OrigLatch, BasicBlock *NewHeader,
77 char LoopRotate::ID = 0;
78 INITIALIZE_PASS_BEGIN(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
79 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
80 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
81 INITIALIZE_PASS_DEPENDENCY(LCSSA)
82 INITIALIZE_PASS_END(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
84 Pass *llvm::createLoopRotatePass() { return new LoopRotate(); }
86 /// Rotate Loop L as many times as possible. Return true if
87 /// the loop is rotated at least once.
88 bool LoopRotate::runOnLoop(Loop *L, LPPassManager &LPM) {
89 LI = &getAnalysis<LoopInfo>();
91 // One loop can be rotated multiple times.
92 bool MadeChange = false;
99 /// Rotate loop LP. Return true if the loop is rotated.
100 bool LoopRotate::rotateLoop(Loop *L) {
101 BasicBlock *OrigPreHeader = L->getLoopPreheader();
102 if (!OrigPreHeader) return false;
104 BasicBlock *OrigLatch = L->getLoopLatch();
105 if (!OrigLatch) return false;
107 BasicBlock *OrigHeader = L->getHeader();
109 // If the loop has only one block then there is not much to rotate.
110 if (L->getBlocks().size() == 1)
113 // If the loop header is not one of the loop exiting blocks then
114 // either this loop is already rotated or it is not
115 // suitable for loop rotation transformations.
116 if (!L->isLoopExiting(OrigHeader))
119 BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator());
122 assert(BI->isConditional() && "Branch Instruction is not conditional");
124 // Updating PHInodes in loops with multiple exits adds complexity.
125 // Keep it simple, and restrict loop rotation to loops with one exit only.
126 // In future, lift this restriction and support for multiple exits if
128 SmallVector<BasicBlock*, 8> ExitBlocks;
129 L->getExitBlocks(ExitBlocks);
130 if (ExitBlocks.size() > 1)
133 // Check size of original header and reject loop if it is very big.
136 Metrics.analyzeBasicBlock(OrigHeader);
137 if (Metrics.NumInsts > MAX_HEADER_SIZE)
141 // Now, this loop is suitable for rotation.
143 // Anything ScalarEvolution may know about this loop or the PHI nodes
144 // in its header will soon be invalidated.
145 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
148 // Find new Loop header. NewHeader is a Header's one and only successor
149 // that is inside loop. Header's other successor is outside the
150 // loop. Otherwise loop is not suitable for rotation.
151 BasicBlock *Exit = BI->getSuccessor(0);
152 BasicBlock *NewHeader = BI->getSuccessor(1);
153 if (L->contains(Exit))
154 std::swap(Exit, NewHeader);
155 assert(NewHeader && "Unable to determine new loop header");
156 assert(L->contains(NewHeader) && !L->contains(Exit) &&
157 "Unable to determine loop header and exit blocks");
159 // This code assumes that the new header has exactly one predecessor.
160 // Remove any single-entry PHI nodes in it.
161 assert(NewHeader->getSinglePredecessor() &&
162 "New header doesn't have one pred!");
163 FoldSingleEntryPHINodes(NewHeader);
165 // Begin by walking OrigHeader and populating ValueMap with an entry for
167 BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
168 ValueToValueMapTy ValueMap;
170 // For PHI nodes, the value available in OldPreHeader is just the
171 // incoming value from OldPreHeader.
172 for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
173 ValueMap[PN] = PN->getIncomingValue(PN->getBasicBlockIndex(OrigPreHeader));
175 // For the rest of the instructions, either hoist to the OrigPreheader if
176 // possible or create a clone in the OldPreHeader if not.
177 TerminatorInst *LoopEntryBranch = OrigPreHeader->getTerminator();
179 Instruction *Inst = I++;
181 // If the instruction's operands are invariant and it doesn't read or write
182 // memory, then it is safe to hoist. Doing this doesn't change the order of
183 // execution in the preheader, but does prevent the instruction from
184 // executing in each iteration of the loop. This means it is safe to hoist
185 // something that might trap, but isn't safe to hoist something that reads
186 // memory (without proving that the loop doesn't write).
187 if (L->hasLoopInvariantOperands(Inst) &&
188 !Inst->mayReadFromMemory() && !Inst->mayWriteToMemory() &&
189 !isa<TerminatorInst>(Inst)) {
190 Inst->moveBefore(LoopEntryBranch);
194 // Otherwise, create a duplicate of the instruction.
195 Instruction *C = Inst->clone();
197 // Eagerly remap the operands of the instruction.
198 RemapInstruction(C, ValueMap,
199 RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
201 // With the operands remapped, see if the instruction constant folds or is
202 // otherwise simplifyable. This commonly occurs because the entry from PHI
203 // nodes allows icmps and other instructions to fold.
204 Value *V = SimplifyInstruction(C);
205 if (V && LI->replacementPreservesLCSSAForm(C, V)) {
206 // If so, then delete the temporary instruction and stick the folded value
211 // Otherwise, stick the new instruction into the new block!
212 C->setName(Inst->getName());
213 C->insertBefore(LoopEntryBranch);
218 // Along with all the other instructions, we just cloned OrigHeader's
219 // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
220 // successors by duplicating their incoming values for OrigHeader.
221 TerminatorInst *TI = OrigHeader->getTerminator();
222 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
223 for (BasicBlock::iterator BI = TI->getSuccessor(i)->begin();
224 PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
225 PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreHeader);
227 // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
228 // OrigPreHeader's old terminator (the original branch into the loop), and
229 // remove the corresponding incoming values from the PHI nodes in OrigHeader.
230 LoopEntryBranch->eraseFromParent();
231 for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I)
232 PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreHeader));
234 // Now fix up users of the instructions in OrigHeader, inserting PHI nodes
237 for (I = OrigHeader->begin(); I != E; ++I) {
238 Value *OrigHeaderVal = I;
239 Value *OrigPreHeaderVal = ValueMap[OrigHeaderVal];
241 // If there are no uses of the value (e.g. because it returns void), there
242 // is nothing to rewrite.
243 if (OrigHeaderVal->use_empty() && OrigPreHeaderVal->use_empty())
246 // The value now exits in two versions: the initial value in the preheader
247 // and the loop "next" value in the original header.
248 SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName());
249 SSA.AddAvailableValue(OrigHeader, OrigHeaderVal);
250 SSA.AddAvailableValue(OrigPreHeader, OrigPreHeaderVal);
252 // Visit each use of the OrigHeader instruction.
253 for (Value::use_iterator UI = OrigHeaderVal->use_begin(),
254 UE = OrigHeaderVal->use_end(); UI != UE; ) {
255 // Grab the use before incrementing the iterator.
256 Use &U = UI.getUse();
258 // Increment the iterator before removing the use from the list.
261 // SSAUpdater can't handle a non-PHI use in the same block as an
262 // earlier def. We can easily handle those cases manually.
263 Instruction *UserInst = cast<Instruction>(U.getUser());
264 if (!isa<PHINode>(UserInst)) {
265 BasicBlock *UserBB = UserInst->getParent();
267 // The original users in the OrigHeader are already using the
268 // original definitions.
269 if (UserBB == OrigHeader)
272 // Users in the OrigPreHeader need to use the value to which the
273 // original definitions are mapped.
274 if (UserBB == OrigPreHeader) {
275 U = OrigPreHeaderVal;
280 // Anything else can be handled by SSAUpdater.
285 // NewHeader is now the header of the loop.
286 L->moveToHeader(NewHeader);
288 // Move the original header to the bottom of the loop, where it now more
289 // naturally belongs. This isn't necessary for correctness, and CodeGen can
290 // usually reorder blocks on its own to fix things like this up, but it's
291 // still nice to keep the IR readable.
293 // The original header should have only one predecessor at this point, since
294 // we checked that the loop had a proper preheader and unique backedge before
296 assert(OrigHeader->getSinglePredecessor() &&
297 "Original loop header has too many predecessors after loop rotation!");
298 OrigHeader->moveAfter(OrigHeader->getSinglePredecessor());
300 // Also, since this original header only has one predecessor, zap its
301 // PHI nodes, which are now trivial.
302 FoldSingleEntryPHINodes(OrigHeader);
304 // TODO: We could just go ahead and merge OrigHeader into its predecessor
305 // at this point, if we don't mind updating dominator info.
307 // Establish a new preheader, update dominators, etc.
308 preserveCanonicalLoopForm(L, OrigHeader, OrigPreHeader, OrigLatch,
316 /// After loop rotation, loop pre-header has multiple sucessors.
317 /// Insert one forwarding basic block to ensure that loop pre-header
318 /// has only one successor.
319 void LoopRotate::preserveCanonicalLoopForm(Loop *L, BasicBlock *OrigHeader,
320 BasicBlock *OrigPreHeader,
321 BasicBlock *OrigLatch,
322 BasicBlock *NewHeader,
325 // Right now original pre-header has two successors, new header and
326 // exit block. Insert new block between original pre-header and
327 // new header such that loop's new pre-header has only one successor.
328 BasicBlock *NewPreHeader =
329 BasicBlock::Create(OrigHeader->getContext(), "bb.nph",
330 OrigHeader->getParent(), NewHeader);
331 LoopInfo &LI = getAnalysis<LoopInfo>();
332 if (Loop *PL = LI.getLoopFor(OrigPreHeader))
333 PL->addBasicBlockToLoop(NewPreHeader, LI.getBase());
334 BranchInst::Create(NewHeader, NewPreHeader);
336 BranchInst *OrigPH_BI = cast<BranchInst>(OrigPreHeader->getTerminator());
337 if (OrigPH_BI->getSuccessor(0) == NewHeader)
338 OrigPH_BI->setSuccessor(0, NewPreHeader);
340 assert(OrigPH_BI->getSuccessor(1) == NewHeader &&
341 "Unexpected original pre-header terminator");
342 OrigPH_BI->setSuccessor(1, NewPreHeader);
346 for (BasicBlock::iterator I = NewHeader->begin();
347 (PN = dyn_cast<PHINode>(I)); ++I) {
348 int index = PN->getBasicBlockIndex(OrigPreHeader);
349 assert(index != -1 && "Expected incoming value from Original PreHeader");
350 PN->setIncomingBlock(index, NewPreHeader);
351 assert(PN->getBasicBlockIndex(OrigPreHeader) == -1 &&
352 "Expected only one incoming value from Original PreHeader");
355 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) {
356 DT->addNewBlock(NewPreHeader, OrigPreHeader);
357 DT->changeImmediateDominator(L->getHeader(), NewPreHeader);
358 DT->changeImmediateDominator(Exit, OrigPreHeader);
359 for (Loop::block_iterator BI = L->block_begin(), BE = L->block_end();
362 if (L->getHeader() != B) {
363 DomTreeNode *Node = DT->getNode(B);
364 if (Node && Node->getBlock() == OrigHeader)
365 DT->changeImmediateDominator(*BI, L->getHeader());
368 DT->changeImmediateDominator(OrigHeader, OrigLatch);
371 if (DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>()) {
372 // New Preheader's dominance frontier is Exit block.
373 DominanceFrontier::DomSetType NewPHSet;
374 NewPHSet.insert(Exit);
375 DF->addBasicBlock(NewPreHeader, NewPHSet);
377 // New Header's dominance frontier now includes itself and Exit block
378 DominanceFrontier::iterator HeadI = DF->find(L->getHeader());
379 if (HeadI != DF->end()) {
380 DominanceFrontier::DomSetType & HeaderSet = HeadI->second;
382 HeaderSet.insert(L->getHeader());
383 HeaderSet.insert(Exit);
385 DominanceFrontier::DomSetType HeaderSet;
386 HeaderSet.insert(L->getHeader());
387 HeaderSet.insert(Exit);
388 DF->addBasicBlock(L->getHeader(), HeaderSet);
391 // Original header (new Loop Latch)'s dominance frontier is Exit.
392 DominanceFrontier::iterator LatchI = DF->find(L->getLoopLatch());
393 if (LatchI != DF->end()) {
394 DominanceFrontier::DomSetType &LatchSet = LatchI->second;
395 LatchSet = LatchI->second;
397 LatchSet.insert(Exit);
399 DominanceFrontier::DomSetType LatchSet;
400 LatchSet.insert(Exit);
401 DF->addBasicBlock(L->getHeader(), LatchSet);
404 // If a loop block dominates new loop latch then add to its frontiers
405 // new header and Exit and remove new latch (which is equal to original
407 BasicBlock *NewLatch = L->getLoopLatch();
409 assert(NewLatch == OrigHeader && "NewLatch is inequal to OrigHeader");
411 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) {
412 for (Loop::block_iterator BI = L->block_begin(), BE = L->block_end();
415 if (!DT->dominates(B, NewLatch)) continue;
417 DominanceFrontier::iterator BDFI = DF->find(B);
418 if (BDFI != DF->end()) {
419 DominanceFrontier::DomSetType &BSet = BDFI->second;
420 BSet.erase(NewLatch);
421 BSet.insert(L->getHeader());
424 DominanceFrontier::DomSetType BSet;
425 BSet.insert(L->getHeader());
427 DF->addBasicBlock(B, BSet);
433 // Preserve canonical loop form, which means Exit block should
434 // have only one predecessor.
435 SplitEdge(L->getLoopLatch(), Exit, this);
437 assert(NewHeader && L->getHeader() == NewHeader &&
438 "Invalid loop header after loop rotation");
439 assert(NewPreHeader && L->getLoopPreheader() == NewPreHeader &&
440 "Invalid loop preheader after loop rotation");
441 assert(L->getLoopLatch() &&
442 "Invalid loop latch after loop rotation");